Predicting the penetration depth during electron beam welding (EBW) is important, but the accuracy of current predictive models is highly varied, depending on the type and number of data used. This paper develops and compares several penetration depth prediction models for EBW and uniquely compares the influence of the number and type of data used, as well as the measurement and modelling methods. Although accelerating voltage, beam current and welding speed data are essential modelling inputs, additional data for beam focal position and beam shape, measured using a novel 4-slit beam probing method, greatly improve the accuracy of predictions for models based on an empirical equation, a second-order regression and an artificial neural network (ANN). Optimised models predict weld depths that deviate, on average, by less than 5% from measured depths, are valid for very broad linear electron beam power density ranges (86–324 J/mm) and are close to the estimated 4% inherent variability in the process and its measurement. Within this linear electron beam power density range, the ANN yields accurate and reliable depth predictions, demanding as few as 36 welding trials, decreasing the number required for models that do not consider beam focal position and shape, for the same targeted accuracy, by more than 60%. Adding large volumes of virtual data generated by less reliable analytical or regression models did not improve the predictive capability for the ANN developed in this study.